Author(s)

Yuan Jiang, Xinchen Xu, Huacheng Cui, Shengyan Song, Zuixing Lin, Zhe Li, He Lin, Xuewen Ding^*

School of Electronic Engineering, Tianjin University of Technology and Education, Tianjin, China.

Time series anomaly detection is important in fields such as industrial control, but faces challenges such as data distribution drifting over time, diverse normal patterns, and training data containing anomalous contamination. In this paper, we propose a time series anomaly detection model RoCA-TFD that integrates trend feature discriminator and expert memory, which introduces a trend feature discriminator (TFD) to recognize the stable, periodic, or abnormal patterns of sequences based on the existing RoCA (Robust Contrastive One-class Anomaly detection) model. Discriminator (TFD) is introduced to identify the stable, periodic, and drifting patterns of sequences, and the Z-score statistical detection combined with the drift detection mechanism of memory comparison is used to realize the timely detection of changes in the data distribution and dynamic prototype update. The model constructs a Mixture-of-Experts (MoE) strategy through lightweight adapters of three patterns, and the patterns discriminated by TFD are routed to weighted fusion of different expert branches. A double-buffered memory system (short-term memory cache + long-term prototype repository) records normal pattern features for assisting drift detection and regulating expert routing. In this study, experiments are conducted on SWaT industrial control database, and the results show that RoCA-TFD improves on Precision, F1 and NAB scores compared to the original RoCA, and achieves more accurate detection of anomalies and fewer false alarms. The method in this paper provides new ideas for time series anomaly detection that includes conceptual drift and multi-modality.

Time Series Anomaly Detection, Concept Drift, Trend Feature Discriminator, Mixture-of-Experts (MoE), Industrial Control Systems

Jiang, Y. , Xu, X. , Cui, H. , Song, S. , Lin, Z. , Li, Z. , Lin, H. and Ding, X. (2025) Time Series Anomaly Detection Based on the Combination of Trend Feature Discrimination and Expert Memory. Journal of Computer and Communications, 13, 65-81. doi: 10.4236/jcc.2025.1310004.